Deep learning based airborne hyperspectral image compression method, device and medium

Abstract

This application discloses an airborne hyperspectral image compression method, device, and medium based on deep learning, relating to the field of deep learning technology. The method involves acquiring a cube of airborne hyperspectral image data containing spatial and spectral dimensions; calculating a reliability index and generating error bound parameters for each spectral band; establishing a reconstruction loop at the encoding end, using the reconstructed data as the prediction context input to a pre-trained deep learning prediction network to generate predicted values ​​and calculate residuals; the pre-trained deep learning prediction network employs a task-driven training approach, incorporating a land cover classification task network and category labels as supervision information, with the training objective consisting of a weighted sum of a rate estimation term, a task loss term, and a reconstruction constraint term; performing near-lossless quantization on the residuals based on the error bound parameters, entropy encoding the quantized residuals to generate a compressed bitstream, and writing the required parameter information for decoding at the bitstream header. This method improves the model's generalization ability through task-driven training, ensuring consistency between encoding and decoding predictions.

Description

Technical Field

[0001] This application relates to the field of deep learning technology, specifically to a deep learning-based airborne hyperspectral image compression method, device, and medium. Background Technology

[0002] Currently, airborne hyperspectral imaging systems can acquire continuous spectral information of ground objects across multiple narrow bands, forming data cubes with high spatial resolution and high spectral dimensionality. These systems are widely used in ground object classification, target identification, anomaly detection, disaster monitoring, and resource surveys. However, due to the large volume of data generated during flight and the high real-time requirements, coupled with limited airborne storage resources and air-to-ground link bandwidth, hyperspectral data typically needs to be compressed on-board before transmission.

[0003] Existing airborne hyperspectral compression methods mainly include predictive coding, transform coding, and standard-based lossless or near-lossless compression schemes. Among them, near-lossless compression methods usually employ a unified error bound or a unified quantization step size to control the maximum error. However, when the spectral noise level, calibration uncertainty, and anomaly rate vary significantly with the band, a unified error bound can lead to insufficient distortion control in key spectral bands or unreasonable overall bit rate configuration.

[0004] On the other hand, deep learning has been applied in image compression in recent years, using neural networks to make predictions and remove data redundancy. However, existing deep learning compression methods mostly focus on pixel reconstruction error as the primary optimization objective. In airborne scenarios, there are still engineering problems such as the consistency of predictions between the encoder and decoder, error controllability, and the adaptation of parameter signaling and computing power constraints. In particular, when the encoder and decoder do not share the original data, how to ensure that both can generate consistent prediction values, so that the residual quantization and reconstruction process can be reproduced, remains a technical challenge.

[0005] Furthermore, the compression objective of airborne hyperspectral data not only focuses on pixel reconstruction quality but also on the performance of downstream application tasks (such as land cover classification and target recognition). Existing compression methods lack optimization mechanisms for downstream tasks, making it difficult to incorporate task constraints into the learning of compression network parameters during the training phase, while maintaining the determinism of the airborne compression process, parameter reliability, and consistency with near-lossless quantization rules during the runtime phase.

[0006] Therefore, many deep learning-based predictive coding methods assume ideal context information during training. However, in actual airborne encoding and decoding processes, there may be differences between the reconstructed data available at the encoding end and the data that the decoding end can obtain. This causes the prediction network to produce inconsistent prediction values ​​at both ends, resulting in error accumulation and decoding drift, which affects compression efficiency and reconstruction quality. Summary of the Invention

[0007] The purpose of this application is to provide an airborne hyperspectral image compression method, computer device, and computer-readable storage medium based on deep learning, in order to solve the problem of inconsistent encoding and decoding prediction in the prior art. When deep learning is introduced for prediction and decorrelation, it is difficult for the encoding and decoding ends to generate prediction values ​​in a consistent manner without sharing the original data, resulting in the inability to reproduce the residual quantization and reconstruction process. When the compression target focuses not only on pixel reconstruction but also on downstream land cover classification and other task indicators, it is difficult to introduce task constraints into the learning of compression network parameters during the training phase, while maintaining the determinism of the airborne compression process, the reliability of parameters, and the consistency of near-lossless quantization rules during the runtime phase.

[0008] To address the aforementioned technical problems, this application provides the following technical solution: a deep learning-based airborne hyperspectral image compression method, comprising: Acquire an airborne hyperspectral image data cube, the data cube comprising spatial and spectral dimensions; For each spectral band, calculate the spectral band reliability index and generate error boundary parameters for the spectral band; At the encoding end, a reconstruction loop is established and the reconstructed data is used as the prediction context input to a pre-trained deep learning prediction network to generate prediction values. The residual between the original data and the prediction values ​​is calculated. The residual is subjected to near-lossless quantization based on the error boundary parameter to generate a quantized residual; The quantization residual is entropy encoded to generate a compressed bitstream, and parameter information for error boundary generation, quantization rules, prediction context configuration, and model identification is written into the bitstream header.

[0009] As an optional scheme of the airborne hyperspectral image compression method based on deep learning described in this application, the spectral band reliability index is constructed by deterministic sources, namely, the estimated value of spectral band noise variance, the spectral band calibration uncertainty, and the spectral band anomaly rate. The spectral anomaly rate represents the proportion of bad elements and strip noise in the spectral band. The estimated value of the spectral noise variance and the spectral calibration uncertainty are obtained from the output of the airborne calibration link or from the statistics of the data cube at the airborne end.

[0010] As an optional scheme of the airborne hyperspectral image compression method based on deep learning described in this application, the error boundary parameter is generated by a deterministic mapping from reliability to error boundary. The mapping includes constraints on the lower limit and upper limit of the error boundary, and the reliability is transformed by a preset scaling factor and a zero-prevention constant before being truncated to between the lower limit and the upper limit. The error bound parameters generated by the mapping are written into the code stream header in the form of spectral segment-level parameters, which are used by the decoding end to quantize and dequantize each spectral segment using a consistent error bound.

[0011] As an optional scheme of the airborne hyperspectral image compression method based on deep learning described in this application, the spectral bands are divided into K spectral band groups, and the division is based on the numerical interval boundary of the spectral band reliability index. For each spectral band group, the group reliability is calculated and a group error boundary parameter is generated. All spectral bands belonging to the same spectral band group are uniformly quantized and dequantized using the group error boundary parameter. The index mapping table of spectral bands to spectral band groups and the group error boundary parameter of each spectral band group are written into the code stream header.

[0012] As an optional scheme of the airborne hyperspectral image compression method based on deep learning described in this application, the near-lossless quantization adopts a quantization step size generation rule that corresponds one-to-one with the error boundary parameter, and the quantization step size generation rule satisfies the following formula:

[0013] in, For spectral band error boundary parameters, Quantization step size for the spectral band; When using the band group error boundary parameter, bands belonging to the same band group share the same quantization step size; The encoding end performs scalar quantization on the residual according to the quantization step size to obtain the quantized residual. The decoding end performs dequantization according to the same quantization step size to reproduce and reconstruct the residual.

[0014] As an optional solution of the airborne hyperspectral image compression method based on deep learning described in this application, wherein: the pre-trained deep learning prediction network uses the prediction context of the reconstruction loop to perform sample-by-sample prediction at the encoding end, and the sample-by-sample prediction processes the spatial pixel position and spectral band position sequentially according to a preset scanning order; The prediction context consists of the reconstructed pixel values ​​in the spatial neighborhood of the current sample to be predicted, the reconstructed spectral values ​​before the current sample to be predicted, and a context mask used to indicate the boundaries of the prediction context. After generating the quantization residual, the encoder immediately performs dequantization and superimposes it with the predicted value to obtain the reconstructed value of the current sample. The reconstructed value is then written into the reconstruction loop buffer for subsequent sample prediction.

[0015] As an optional scheme of the airborne hyperspectral image compression method based on deep learning described in this application, the parameter information written in the code stream header includes: spectral band-level error boundary parameters or spectral band group error boundary parameters, index mapping table from spectral band to spectral band group, lower and upper error boundaries, proportional coefficient and zero-prevention constant used for reliability-to-error boundary mapping, quantization step size generation rule identifier, model identifier of pre-trained deep learning prediction network, prediction context configuration parameters, and scanning order identifier, wherein the prediction context configuration parameters include the spatial neighborhood window size and the number of spectral reference bands; After reading the parameter information from the bitstream header, the decoding end performs the following decoding process: entropy decoding is performed on the entropy-encoded data to obtain the quantization residual; The quantization step size of each spectral segment is calculated based on the error boundary parameters and quantization step size generation rules in the code stream header, and inverse quantization is performed to obtain the reconstruction residual; At the decoding end, a reconstruction loop buffer consistent with that at the encoding end is established. A prediction context is generated according to the scanning order indicated by the code stream header and the prediction context configuration parameters. The prediction context is input into the pre-trained deep learning prediction network corresponding to the model identifier to generate prediction values. The prediction values ​​are superimposed with the reconstruction residuals to obtain reconstruction data. At the same time, the reconstruction data is written into the reconstruction loop buffer for subsequent sample prediction calls.

[0016] As an optional scheme of the airborne hyperspectral image compression method based on deep learning described in this application, the pre-trained deep learning prediction network obtains network parameters through task-driven training during the offline training phase. During the training phase, a land cover classification task network is introduced, and the category labels of the training samples are used as supervision information. The training objective consists of a bitrate estimation term, a task loss term, and a reconstruction constraint term, which are weighted and summed using preset non-negative weight coefficients. The task loss term takes the form of cross-entropy and satisfies the following equation:

[0017] in, This is the task loss value. For the number of categories, The first label for the category dimensional components, The first output result of the task network on the reconstructed data dimensional components, This is a logarithmic operation.

[0018] This application provides a deep learning-based airborne hyperspectral image compression system, including: The acquisition module is used to acquire an airborne hyperspectral image data cube, wherein the data cube contains spatial dimension and spectral dimension; The calculation module is used to calculate the reliability index of each spectral band and generate error boundary parameters for the spectral band. The prediction module is used to establish a reconstruction loop at the encoding end and use the reconstructed data as the prediction context input to the pre-trained deep learning prediction network to generate predicted values, and calculate the residual between the original data and the predicted values. The pre-trained deep learning prediction network obtains network parameters in the offline training stage using a task-driven training method. In the training stage, a land cover classification task network is introduced and the category labels of the training samples are used as supervision information. The training objective consists of a bit rate estimation term, a task loss term, and a reconstruction constraint term, and is weighted and summed using preset non-negative weight coefficients. The generation module is used to perform near-lossless quantization on the residual based on the error boundary parameter to generate quantized residual; The writing module is used to entropy encode the quantization residual to generate a compressed bitstream, and write parameter information for error boundary generation, quantization rules, prediction context configuration and model identification at the decoding end into the bitstream header.

[0019] This application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the aforementioned deep learning-based airborne hyperspectral image compression method.

[0020] This application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned deep learning-based airborne hyperspectral image compression method.

[0021] The beneficial effects of this application are as follows: By constructing a reliability index at the spectral band level and generating error bound parameters accordingly, near-lossless quantization can implement differentiated error constraints based on spectral band reliability. Simultaneously, by establishing a reconstruction loop at the encoding end and using the reconstructed context as input to a pre-trained deep learning prediction network, the encoding and decoding ends can reproduce predicted values ​​under the same model identifier and context configuration, thereby ensuring the consistency of residual quantization, inverse quantization, and reconstruction processes. By establishing a reconstruction loop at the encoding end and using the reconstructed data as the prediction context, it is ensured that the encoding and decoding ends use the same reference information, thus solving the aforementioned inconsistency problem.

[0022] Furthermore, this application introduces a task network during the offline training phase and uses task loss to participate in the joint optimization of the compression network, so that the predicted network parameters obtained from training are combined with task-related constraints. During the inference phase, the airborne terminal only calls the trained predicted network and entropy coding module and transmits parameter information such as error bounds, group mapping, context configuration and model identification through the code stream header signaling, thus forming a hyperspectral compression process and an achievable code stream description method for airborne scenarios. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating the overall process of an airborne hyperspectral image compression method based on deep learning, provided in one embodiment of this application. Detailed Implementation

[0025] To make this application more apparent and understandable, the specific embodiments of this application are described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this application.

[0026] The technical problems addressed in this application are: ① How to avoid insufficient distortion control in key spectral bands or unreasonable overall bitrate configuration when the spectral noise level, calibration uncertainty, and anomaly rate vary significantly with the band. ② How to ensure that the encoder and decoder can consistently generate predicted values ​​without sharing the original data when introducing deep learning for prediction and decorrelation, thereby making the residual quantization and reconstruction process reproducible. ③ How to incorporate task constraints into the learning of compression network parameters during the training phase, while maintaining the determinism, parameter reliability, and consistency of near-lossless quantization rules of the airborne compression process during the runtime phase, when the compression objective focuses not only on pixel reconstruction but also on downstream land cover classification and other task indicators.

[0027] It should be noted that the terms used in the following embodiments are defined as follows: The spectral reliability index is a comprehensive evaluation index that integrates the noise variance estimate, calibration uncertainty, and anomaly rate. The larger the value, the higher the quality of the spectral band. The error bound parameter refers to the maximum allowable reconstruction error in near-lossless quantization, and its unit is grayscale value; Quantization step size refers to the discretization interval during residual quantization. Conversion efficiency refers to the ratio of the output data volume to the input data volume; Prediction context refers to the range of reconstructed data used by a pre-trained deep learning prediction network when making predictions.

[0028] Example 1, referring to Figure 1This is one embodiment of the present application, which provides an airborne hyperspectral image compression method based on deep learning, including five main steps from S100 to S500. Detailed steps include: S100: Acquire an airborne hyperspectral image data cube, the data cube containing spatial and spectral dimensions.

[0029] S200: Calculates the reliability index for each spectral band and generates error boundary parameters for the spectral band.

[0030] S300: Establish a reconstruction loop at the encoding end and use the reconstructed data as the prediction context input to the pre-trained deep learning prediction network to generate predicted values, and calculate the residual between the original data and the predicted values.

[0031] S400: Perform near-lossless quantization on the residual based on the error boundary parameter to generate quantized residual.

[0032] S500: The quantization residual is entropy encoded to generate a compressed bitstream, and parameter information for error boundary generation, quantization rules, prediction context configuration and model identification is written into the bitstream header.

[0033] Example 2, an embodiment of this application, provides a deep learning-based airborne hyperspectral image compression method based on the previous embodiment, including: S100: Acquire an airborne hyperspectral image data cube, the data cube containing spatial and spectral dimensions.

[0034] S200: Calculates the reliability index for each spectral band and generates error boundary parameters for the spectral band.

[0035] S201: The spectral band reliability index is constructed from deterministic sources, namely, the estimated value of spectral band noise variance, the spectral band calibration uncertainty, and the spectral band anomaly rate; The spectral anomaly rate represents the proportion of bad elements and strip noise in the spectral band. The estimated value of the spectral noise variance and the spectral calibration uncertainty are obtained from the output of the airborne calibration link or from the statistics of the data cube at the airborne end.

[0036] S202: Further, the error boundary parameter is generated through a deterministic mapping from reliability to error boundary. The mapping includes constraints on the lower limit and the upper limit of the error boundary. The reliability is transformed by a preset proportional coefficient and a zero-prevention constant and then truncated to the area between the lower limit and the upper limit. The error bound parameters generated by the mapping are written into the code stream header in the form of spectral segment-level parameters, which are used by the decoding end to quantize and dequantize each spectral segment using a consistent error bound.

[0037] It should be noted that the preset proportional coefficient is used to control the conversion scale from reliability index to error boundary parameter, and it is obtained as follows: After calculating the reliability index for all spectral segments of the data cube, the reliability index for each spectral segment is extracted to form a reliability vector. Statistical operations are then performed on the reliability vector to obtain the reliability center value. Extract the lower limit and upper limit of the error boundary, and calculate their arithmetic mean as the center value of the error boundary; The preset proportional coefficient is obtained by multiplying the reliability center value and the error boundary center value.

[0038] The reliability center value is the arithmetic mean of the reliability indices for all spectrum bands.

[0039] Once the preset scaling factor is determined, when performing the mapping from reliability to error bound for each spectral band at the encoding end, the preset scaling factor is used as the scaling factor for the inverse proportional transformation, and the mapping operation is completed in combination with the zero-prevention constant and the truncation boundary.

[0040] In this embodiment, based on experience, when the lower limit of the error threshold is 1, the upper limit of the error threshold is 5, and the reliability center value is 1.5, the preset proportional coefficient is 3. S203: Divide the spectrum into K spectrum groups based on the numerical range boundary of the spectrum reliability index. Calculate the group reliability for each spectrum group and generate a group error boundary parameter. Perform quantization and dequantization on all spectrum groups belonging to the same spectrum group using the group error boundary parameter. Write the index mapping table of spectrum groups to spectrum groups and the group error boundary parameter of each spectrum group into the code stream header.

[0041] It should be noted that the reliability index of the spectral band is calculated according to the following formula:

[0042] in, This is the estimated variance of the spectral noise. For the calibration uncertainty of the spectral band, For spectral anomaly rate, These are preset non-negative weighting coefficients.

[0043] It should be noted that the preset non-negative weighting coefficient is used to weight and combine the three deterministic sources: the estimated variance of the spectral band noise, the spectral band calibration uncertainty, and the spectral band anomaly rate. The determination method is as follows: Based on the noise characteristics of the airborne hyperspectral imaging system, determine the weighting coefficients of the estimated spectral noise variance. Based on the calibration accuracy of the airborne calibration link, determine the weighting coefficients for the spectral calibration uncertainty; The weighting coefficients for the spectral anomaly rate are determined based on the detection sensitivity of spectral band bad elements and strip noise.

[0044] In the calculation of the spectral band reliability index, each weight coefficient is multiplied by its corresponding deterministic source and then summed. The reciprocal of the sum is taken to obtain the spectral band reliability index. The preset non-negative weight coefficients do not need to be written into the code stream header, and the decoding end does not need to reproduce the calculation process of the spectral band reliability index.

[0045] In this embodiment, the weighting coefficient of the spectral noise variance estimate is 0.5, the weighting coefficient of the spectral calibration uncertainty is 0.3, and the weighting coefficient of the spectral anomaly rate is 0.2.

[0046] S300: Establish a reconstruction loop at the encoding end and use the reconstructed data as the prediction context input to the pre-trained deep learning prediction network to generate predicted values, and calculate the residual between the original data and the predicted values.

[0047] S301: The pre-trained deep learning prediction network uses the prediction context of the reconstruction loop at the encoding end to perform sample-by-sample prediction. The sample-by-sample prediction processes the spatial pixel position and spectral position sequentially according to a preset scanning order.

[0048] The prediction context consists of the reconstructed pixel values ​​in the spatial neighborhood of the current sample to be predicted, the reconstructed spectral values ​​before the current sample to be predicted, and a context mask used to indicate the boundaries of the prediction context.

[0049] It should be noted that the preset scanning order is used to guide the traversal order of the encoding and decoding ends in processing the spatial pixel positions and spectral band positions in the data cube sample by sample. The method for determining this order is as follows: A spectral segment-first nested scanning strategy is adopted for the data cube. For each spatial pixel position, all spectral segments at that position are processed in ascending order of spectral segment index. After completion, the process moves to the next spatial pixel position to continue processing.

[0050] The spatial pixel positions are traversed using a raster scan order, starting from the first row and first column of the data cube, and proceeding pixel by pixel along the row direction to the end of the row, then moving to the next row to continue scanning, until all spatial positions have been processed.

[0051] The spectral priority scanning strategy enables the prediction context of the current sample to be predicted to include the reconstructed spectral values ​​at the same spatial location, utilizes the correlation between spectral segments to generate predicted values, reduces residual amplitude, and improves entropy coding efficiency.

[0052] The scan order identifier is written to the code stream header. After the decoder reads it, it performs sample-by-sample decoding in the same scan order as the encoder to ensure that the update order of the reconstruction loop buffer is synchronized.

[0053] S302: After generating the quantization residual, the encoder immediately performs dequantization and superimposes it with the predicted value to obtain the reconstructed value of the current sample. The reconstructed value is then written into the reconstruction loop buffer for subsequent sample prediction.

[0054] S400: Perform near-lossless quantization on the residual based on the error boundary parameter to generate quantized residual.

[0055] S401: The near-lossless quantization adopts a quantization step size generation rule that corresponds one-to-one with the error boundary parameter, and the quantization step size generation rule satisfies the following formula:

[0056] in, For spectral band error boundary parameters, Quantization step size for the spectral band; When using the band group error boundary parameter, bands belonging to the same band group share the same quantization step size.

[0057] S402: The encoding end performs scalar quantization on the residual according to the quantization step size to obtain the quantized residual. The decoding end performs dequantization according to the same quantization step size to reproduce and reconstruct the residual.

[0058] S500: The quantization residual is entropy encoded to generate a compressed bitstream, and parameter information for error boundary generation, quantization rules, prediction context configuration and model identification is written into the bitstream header.

[0059] S501: The parameter information written in the code stream header includes: spectral segment-level error boundary parameters or spectral segment group error boundary parameters, index mapping table from spectral segment to spectral segment group, lower and upper error boundaries, proportional coefficient and zero-prevention constant used for reliability to error boundary mapping, quantization step size generation rule identifier, model identifier of pre-trained deep learning prediction network, prediction context configuration parameters, and scan order identifier, wherein the prediction context configuration parameters include the spatial neighborhood window size and the number of spectral reference segments; S502: After reading the parameter information from the bitstream header, the decoding end performs the following decoding process: entropy decoding is performed on the entropy encoded data to obtain the quantization residual; The quantization step size for each spectral segment is calculated based on the error boundary parameters and quantization step size generation rules in the code stream header, and inverse quantization is performed to obtain the reconstruction residual.

[0060] S503: Establish a reconstruction loop buffer at the decoding end that is consistent with that at the encoding end. Generate a prediction context according to the scanning order indicated by the code stream header and the prediction context configuration parameters. Input the prediction context into the pre-trained deep learning prediction network corresponding to the model identifier to generate prediction values. Superimpose the prediction values ​​with the reconstruction residuals to obtain reconstruction data. At the same time, write the reconstruction data into the reconstruction loop buffer for subsequent sample prediction calls.

[0061] S504: The pre-trained deep learning prediction network obtains network parameters through task-driven training during the offline training phase. A land cover classification task network is introduced during training, and the category labels of training samples are used as supervision information. The training objective consists of a bitrate estimation term, a task loss term, and a reconstruction constraint term, which are weighted and summed using preset non-negative weight coefficients. The task loss term takes the form of cross-entropy and satisfies the following formula:

[0062] in, For the number of categories, The first label for the category dimensional components, The first output result of the task network on the reconstructed data Dimensional components. During the inference phase, the pre-trained deep learning prediction network, which has already been trained, is invoked on the onboard device to perform all the aforementioned steps without invoking the task network.

[0063] It should be noted that the preset non-negative weight coefficients of the bitrate estimation term, task loss term, and reconstruction constraint term in the training objective are determined based on the balance requirements between compression performance and task performance.

[0064] The weight coefficients of the bitrate estimation term are determined based on the constraint strength of the target compression ratio. Increasing the weight coefficients of the bitrate estimation term makes the training process more inclined to reduce the compression bitrate. The weight coefficients of the task loss term are determined based on the accuracy requirements of the land cover classification task. Increasing the weight coefficients of the task loss term makes the training process more inclined to retain spectral features that are beneficial to classification. The weight coefficients of the reconstruction constraint terms are determined based on the fidelity requirements of the reconstructed data. Increasing the weight coefficients of the reconstruction constraint terms makes the training process more inclined to reduce reconstruction error.

[0065] In this embodiment, based on experience, the weight coefficient of the bitrate estimation term is set to 0.4, the weight coefficient of the task loss term is set to 0.4, and the weight coefficient of the reconstruction constraint term is set to 0.2.

[0066] Example 3 See Figure 1 This embodiment provides a deep learning-based airborne hyperspectral image compression method, applied to the data compression scenario of an airborne hyperspectral imaging system. The method includes a complete process such as data acquisition, reliability calculation, error bound generation, reconstruction loop establishment, deep learning prediction, near-lossless quantization, and entropy coding.

[0067] First, step S100 is executed to acquire the airborne hyperspectral image data cube. The airborne hyperspectral image data cube is formed by acquiring continuous spectral information of ground objects across multiple narrow bands using an airborne hyperspectral imaging system. It comprises two dimensions: spatial and spectral. The spatial dimension refers to the pixel distribution of the data cube in the horizontal and vertical directions, while the spectral dimension refers to the number of spectral channels corresponding to different bands. The data format of the airborne hyperspectral image data cube is a three-dimensional array. The first and second dimensions represent spatial location, and the third dimension represents spectral location. The value of each data element represents the radiant intensity at the corresponding spatial and spectral location, typically represented using 8-bit or 12-bit quantization.

[0068] Next, step S200 is executed to calculate the spectral band reliability index for each spectral band and generate error bound parameters for the spectral band. The spectral band reliability index is constructed from deterministic sources, namely the estimated spectral band noise variance, the spectral band calibration uncertainty, and the spectral band anomaly rate. The estimated spectral band noise variance reflects the noise level of the spectral band; a larger value indicates stronger noise. The spectral band calibration uncertainty reflects the calibration accuracy; a larger value indicates a larger calibration error. The spectral band anomaly rate characterizes the proportion of bad elements and strip noise in the spectral band; a larger value indicates more anomalous data. The estimated spectral band noise variance and the spectral band calibration uncertainty are obtained from the output of the airborne calibration link or from the statistics of the data cube calculated by the airborne terminal.

[0069] The spectral reliability index is calculated using the following formula:

[0070] in, This is the estimated variance of the spectral noise. For the calibration uncertainty of the spectral band, For spectral anomaly rate, , , Preset non-negative weighting coefficients are used to weight the three deterministic sources, determined based on the noise characteristics, calibration accuracy, and anomaly detection sensitivity of the airborne hyperspectral imaging system. In calculating the spectral reliability index, each weighting coefficient is multiplied by its corresponding deterministic source, and the sum is taken as the reciprocal of the sum to obtain the spectral reliability index. A higher spectral reliability index indicates higher spectral quality and a smaller allowable compression error; a lower spectral reliability index indicates lower spectral quality and a larger allowable compression error. In this embodiment, the weighting coefficient for the spectral noise variance estimate is 0.5, the weighting coefficient for the spectral calibration uncertainty is 0.3, and the weighting coefficient for the spectral anomaly rate is 0.2.

[0071] Error bound parameters are generated through a deterministic mapping from reliability to error bounds. This deterministic mapping includes constraints on the lower and upper bounds of the error bounds, and transforms the reliability using a preset scaling factor and a zero-prevention constant before truncating it to between the lower and upper bounds. The preset scaling factor controls the conversion scale from reliability indices to error bound parameters. It is obtained as follows: after calculating reliability indices for all spectral segments of the data cube, the reliability indices for each segment are extracted to form a reliability vector. Statistical operations are performed on the reliability vector to obtain the reliability center value. The lower and upper bounds of the error bounds are extracted, and their arithmetic mean is calculated as the error bound center value. The reliability center value and the error bound center value are multiplied to obtain the preset scaling factor. The reliability center value is the arithmetic mean of the reliability indices for all spectral segments. After the preset scaling factor is determined, when performing the reliability-to-error bound mapping for each spectral segment at the encoding end, the preset scaling factor is used as the scaling factor for the inverse proportional transformation, combined with the zero-prevention constant and the truncation boundary to complete the mapping operation. In this embodiment, when the lower limit of the error boundary is 1, the upper limit of the error boundary is 5, and the reliability center value is 1.5, the preset scaling factor is 3. The error boundary parameters generated by mapping are written into the code stream header in the form of spectral segment-level parameters, which are used by the decoding end to quantize and dequantize each spectral segment using a consistent error boundary.

[0072] In some embodiments, the spectral bands are divided into K spectral band groups, based on the numerical range boundaries of the spectral band reliability index. For each spectral band group, the group reliability is calculated and a group error boundary parameter is generated. All spectral bands belonging to the same group are uniformly quantized and dequantized using this group error boundary parameter. An index mapping table of spectral bands to spectral band groups and the group error boundary parameter for each spectral band group are written into the bitstream header. The advantage of the spectral band grouping strategy is reduced parameter signaling overhead. When the number of spectral bands is large, using the spectral band group error boundary parameter can reduce the number of parameters that need to be transmitted in the bitstream header, thereby improving the overall compression efficiency.

[0073] Next, step S300 is executed, establishing a reconstruction loop at the encoding end and using the reconstructed data as the prediction context input to the pre-trained deep learning prediction network to generate predicted values, calculating the residual between the original data and the predicted values. The reconstruction loop refers to simulating the reconstruction process at the decoding end at the encoding end, ensuring that the encoding and decoding ends perform subsequent processing based on the same reconstructed data, thereby guaranteeing consistency between encoding and decoding predictions. The pre-trained deep learning prediction network performs sample-by-sample prediction at the encoding end using the prediction context of the reconstruction loop, processing spatial pixel positions and spectral band positions sequentially according to a preset scanning order.

[0074] The preset scanning order guides the traversal order of the encoder and decoder for processing spatial pixel and spectral segment positions in the data cube sample by sample. A spectral segment-first nested scanning strategy is adopted for the data cube. For each spatial pixel position, all spectral segments at that position are processed sequentially according to the ascending spectral segment index. After completion, the process moves to the next spatial pixel position. The traversal of spatial pixel positions adopts a raster scan order, starting from the first row and first column of the data cube, advancing pixel by pixel along the row direction to the end of the row, then moving to the next row to continue scanning until all spatial positions are processed. The spectral segment-first scanning strategy ensures that the prediction context of the current sample to be predicted can include the reconstructed spectral segment values ​​at the same spatial position. It utilizes the correlation between spectral segments to generate predicted values, reducing residual amplitude and improving entropy coding efficiency. The scanning order identifier is written to the bitstream header. After reading it, the decoder performs sample-by-sample decoding according to the same scanning order as the encoder, ensuring the synchronization of the update order of the reconstruction loop buffer.

[0075] The prediction context consists of reconstructed pixel values ​​within the spatial neighborhood of the current sample to be predicted, reconstructed spectral segment values ​​preceding the current sample, and a context mask indicating the boundaries of the prediction context. Reconstructed pixel values ​​within the spatial neighborhood refer to the reconstructed data of the spatially adjacent positions of the current sample to be predicted, typically including pixel values ​​at positions such as above, left, upper left, and upper right. Reconstructed spectral segment values ​​preceding the current sample refer to the reconstructed data in the spectral segment dimension, following an ascending order of spectral segment indices; all spectral segments preceding the current segment have been reconstructed. The context mask indicates the effective range of the prediction context. At the boundary positions of the data cube, some spatial neighborhood positions or spectral segment positions lack reconstructed data. In these cases, the context mask marks the effective and ineffective positions, and the pre-trained deep learning prediction network only uses data from the effective positions for prediction.

[0076] A pre-trained deep learning prediction network receives the prediction context as input and outputs the predicted value for the current sample. The structure of the pre-trained deep learning prediction network includes basic units such as convolutional layers, activation layers, and normalization layers. Through this multi-layered network structure, it extracts spatial and spectral correlations from the prediction context to generate a prediction for the current sample. The parameters of the pre-trained deep learning prediction network are obtained through an offline training phase. After training, the network parameters are fixed, and the same network parameters are used for prediction at both the encoding and decoding ends.

[0077] Calculate the residual between the original data and the predicted value. The residual equals the original data minus the predicted value. The residual represents the difference between the original data and the predicted value. The smaller the magnitude of the residual, the more accurate the prediction and the higher the compression efficiency.

[0078] After generating the quantization residual, the encoder immediately performs dequantization and superimposes it with the predicted value to obtain the reconstructed value of the current sample. The reconstructed value is written to the reconstruction loop buffer for subsequent sample prediction. The reconstruction loop buffer stores the reconstructed data and is updated step by step according to the scan order to ensure that the encoder and decoder perform predictions based on the same reconstructed data, thereby guaranteeing the consistency of the predicted values.

[0079] Next, step S400 is executed, whereby near-lossless quantization is performed on the residuals based on the error bounds parameters to generate quantization residuals. Near-lossless quantization employs a quantization step size generation rule that corresponds one-to-one with the error bounds parameters, and the quantization step size generation rule satisfies the following formula:

[0080] in, For spectral band error boundary parameters, This refers to the quantization step size for a spectral band. When using the spectral band group error bound parameter, spectral bands belonging to the same group share the same quantization step size. The choice of quantization step size is based on the error bound parameter. A larger error bound parameter results in a larger quantization step size, a sparser range of residual values ​​after quantization, and higher entropy coding efficiency, but also a larger reconstruction error. Conversely, a smaller error bound parameter results in a smaller quantization step size, a denser range of residual values ​​after quantization, and smaller reconstruction error, but lower entropy coding efficiency. Through the above formula, the quantization step size and the error bound parameter form a linear relationship, ensuring that the reconstruction error after quantization does not exceed the error bound parameter.

[0081] The encoder performs scalar quantization on the residuals, rounding to the nearest integer based on the quantization step size, to obtain the quantized residuals. Scalar quantization involves quantizing each residual value independently. The process is as follows: divide the residual value by the quantization step size, and round the quotient to the nearest integer. The quantized residuals are integers, and their range is affected by the quantization step size; the larger the quantization step size, the smaller the range of values ​​for the quantized residuals. The decoder performs dequantization with the same quantization step size to reproduce the reconstructed residuals. The dequantization process is as follows: multiply the quantized residuals by the quantization step size to obtain the reconstructed residuals. The error between the reconstructed residuals and the original residuals is controlled by the quantization step size, with a maximum error not exceeding half the quantization step size, i.e., not exceeding the error boundary parameter.

[0082] Next, step S500 is executed to perform entropy encoding on the quantization residuals to generate a compressed bitstream. Parameter information for generating error bounds, quantization rules, prediction context configuration, and model identification for the decoding end is written into the bitstream header. Entropy encoding refers to lossless encoding based on the probability distribution of the quantization residuals. Commonly used entropy encoding methods include arithmetic coding and Huffman coding. Entropy encoding utilizes the statistical characteristics of the quantization residuals, assigning shorter codewords to symbols with high occurrence probabilities and longer codewords to symbols with low occurrence probabilities, thereby achieving data compression. The compressed bitstream generated after entropy encoding consists of two parts: a bitstream header and bitstream data. The bitstream header stores the parameter information required for the decoding end to reproduce the compression process, and the bitstream data stores the entropy-encoded quantization residuals.

[0083] The parameters written in the bitstream header include: spectral segment-level error bound parameters or spectral segment group error bound parameters, an index mapping table from spectral segment to spectral segment group, lower and upper error bounds, the scaling factor and zero-prevention constant used for the reliability-to-error bound mapping, quantization step size generation rule identifier, model identifier of the pre-trained deep learning prediction network, prediction context configuration parameters, and scan order identifier. The prediction context configuration parameters include the spatial neighborhood window size and the number of spectral reference segments. The spatial neighborhood window size refers to the spatial neighborhood range included in the prediction context, and the number of spectral reference segments refers to the number of spectral segments included in the prediction context. The model identifier indicates the specific model version of the pre-trained deep learning prediction network; the decoder loads the corresponding network parameters based on the model identifier.

[0084] After reading parameter information from the bitstream header, the decoder performs the following decoding process: First, entropy decoding is performed on the entropy-encoded data to obtain the quantization residual. The entropy decoding process is the inverse of the entropy encoding process. The quantization residual is reproduced based on the bitstream data and entropy encoding rules. Next, the quantization step size for each spectral segment is calculated based on the error bound parameters and quantization step size generation rules in the bitstream header, and inverse quantization is performed to obtain the reconstruction residual. The inverse quantization process involves multiplying the quantization residual by the quantization step size to obtain the reconstruction residual. Then, a reconstruction loop buffer consistent with that of the encoder is established at the decoder. A prediction context is generated according to the scanning order indicated by the bitstream header and the prediction context configuration parameters. The prediction context is input into the pre-trained deep learning prediction network corresponding to the model identifier to generate prediction values. The prediction values ​​are then superimposed with the reconstruction residual to obtain the reconstruction data. Simultaneously, the reconstruction data is written to the reconstruction loop buffer for subsequent sample prediction. The reconstruction loop buffer at the decoder and the reconstruction loop buffer at the encoder are updated sample by sample under the same scanning order and prediction context configuration, ensuring that the prediction contexts of the encoder and decoder are completely consistent, thereby guaranteeing the consistency of the prediction values ​​and the reproducibility of the reconstruction process.

[0085] Through the above methods, this embodiment achieves near-lossless quantization based on dynamic error bounds of spectral band reliability, avoiding the problems of insufficient control of key spectral band distortion or unreasonable overall bitrate configuration caused by unified error bounds. By establishing a reconstruction loop at the encoding end and using reconstructed data as the prediction context, it ensures that the encoding and decoding ends can reproduce the predicted values ​​under the same model identifier and the same context configuration, ensuring the consistency of residual quantization, inverse quantization and reconstruction processes, and eliminating error accumulation caused by prediction drift. By writing complete parameter information into the bitstream header, a hyperspectral compression process and an achievable bitstream description method for airborne scenarios are formed.

[0086] Example 4: The difference from Example 3 is that this example further illustrates the process of obtaining network parameters by using a task-driven training method during the offline training phase of a pre-trained deep learning prediction network.

[0087] The pre-trained deep learning prediction network obtains its parameters through task-driven training during the offline training phase. A land cover classification task network is introduced during training, using the class labels of the training samples as supervision information. The land cover classification task network is a neural network specifically designed for classifying land covers in hyperspectral images; its input is hyperspectral image data, and its output is the probability distribution of each class. The class labels of the training samples refer to the true class labels corresponding to each sample in the training data, used to supervise the training of the land cover classification task network.

[0088] The training objective consists of a bitrate estimation term, a task loss term, and a reconstruction constraint term, which are weighted and summed using preset non-negative weight coefficients. The bitrate estimation term estimates the compressed bitrate, encouraging the network to generate sparser quantization residuals, thereby reducing the bitrate after entropy encoding. The task loss term measures the performance of the reconstructed data on the land cover classification task, encouraging the network to retain spectral features beneficial for classification. The reconstruction constraint term measures the difference between the reconstructed data and the original data, encouraging the network to reduce reconstruction errors.

[0089] The task loss term takes the form of cross-entropy and satisfies the following formula:

[0090] Where C is the number of categories, The c-th dimension component labeled for the category, Let be the c-th dimension of the task network's output on the reconstructed data. Cross-entropy loss measures the difference between the class labels and the task network's output; a smaller value indicates better classification performance.

[0091] The preset non-negative weight coefficients of the bitrate estimation term, task loss term, and reconstruction constraint term in the training objective are determined based on the balance requirements between compression performance and task performance. The weight coefficient of the bitrate estimation term is determined based on the constraint strength of the target compression ratio; increasing the weight coefficient of the bitrate estimation term makes the training process more inclined to reduce the compression bitrate. The weight coefficient of the task loss term is determined based on the accuracy requirements of the land cover classification task; increasing the weight coefficient of the task loss term makes the training process more inclined to preserve spectral features beneficial to classification. The weight coefficient of the reconstruction constraint term is determined based on the fidelity requirements of the reconstructed data; increasing the weight coefficient of the reconstruction constraint term makes the training process more inclined to reduce reconstruction error. In this embodiment, the weight coefficient of the bitrate estimation term is 0.4, the weight coefficient of the task loss term is 0.4, and the weight coefficient of the reconstruction constraint term is 0.2.

[0092] During the inference phase, the pre-trained deep learning prediction network, which has already been trained, is invoked on the airborne end to perform all the steps described in Example 1 without invoking the task network. The task network is only used during the training phase to provide supervision information for the task loss. After training is completed, the parameters of the pre-trained deep learning prediction network already contain task-related optimization information, so there is no need to invoke the task network again during the inference phase, thereby maintaining the determinism and parameter reliability of the airborne compression process.

[0093] In this embodiment, the pre-trained deep learning prediction network incorporates the constraints of the land cover classification task during the training phase, combining the compression network parameters with the requirements of downstream tasks. While maintaining near-lossless error constraints, it improves the performance of tasks such as land cover classification of reconstructed data, realizing a hyperspectral compression process for airborne scenarios and compression parameter configuration for downstream tasks.

[0094] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0095] In response to the above problems, the beneficial effects of this application include at least the following: (1) By constructing a reliability index at the spectral band level and generating differentiated error bound parameters accordingly, near-lossless quantization can implement differentiated error constraints according to the reliability of the spectral band, thus avoiding the problem of insufficient control of key spectral band distortion or unreasonable overall bit rate configuration caused by a unified error bound. (2) By establishing a reconstruction loop at the encoding end and using the reconstructed data as the prediction context input to the pre-trained deep learning prediction network, the encoding end and the decoding end can reproduce the prediction value under the same model identifier and the same context configuration, which ensures the consistency of residual quantization, dequantization and reconstruction process and eliminates the error accumulation caused by prediction drift. (3) By introducing a land cover classification task network in the offline training stage and using task loss to participate in the joint optimization of the compression network, the predicted network parameters obtained from the training are combined with the downstream task requirements, which improves the performance of land cover classification and other tasks of the reconstructed data while maintaining the determinism of the airborne compression process. (4) By writing error boundary parameters, group mapping, context configuration and model identification and other parameter information into the code stream header, a complete parameter signaling mechanism is formed to ensure that the decoding end can correctly reproduce the encoding process, and realize the hyperspectral compression process for airborne scenarios and the feasible code stream description method.

[0096] Example 5: This application provides an airborne hyperspectral image compression system based on deep learning, implementing the airborne hyperspectral image compression method based on deep learning proposed in the above embodiments, including: The acquisition module is used to acquire an airborne hyperspectral image data cube, wherein the data cube contains spatial dimension and spectral dimension; The calculation module is used to calculate the reliability index of each spectral band and generate error boundary parameters for the spectral band. The prediction module is used to establish a reconstruction loop at the encoding end and use the reconstructed data as the prediction context input to the pre-trained deep learning prediction network to generate predicted values, and calculate the residual between the original data and the predicted values. The pre-trained deep learning prediction network obtains network parameters in the offline training stage using a task-driven training method. In the training stage, a land cover classification task network is introduced and the category labels of the training samples are used as supervision information. The training objective consists of a bit rate estimation term, a task loss term, and a reconstruction constraint term, and is weighted and summed using preset non-negative weight coefficients. The generation module is used to perform near-lossless quantization on the residual based on the error boundary parameter to generate quantized residual; The writing module is used to entropy encode the quantization residual to generate a compressed bitstream, and write parameter information for error boundary generation, quantization rules, prediction context configuration and model identification at the decoding end into the bitstream header.

[0097] Example 6: This example also provides an electronic device applicable to the airborne hyperspectral image compression method based on deep learning, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the airborne hyperspectral image compression method based on deep learning as proposed in the above examples.

[0098] Example 7: This example also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements the airborne hyperspectral image compression method based on deep learning as proposed in the above examples.

[0099] The storage medium proposed in this embodiment and the airborne hyperspectral image compression method based on deep learning proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0100] Based on the above description of the implementation methods, those skilled in the art can clearly understand that this application can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of this application.

[0101] It should be noted that the above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit it. Although this application has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of this application without departing from the spirit and scope of the technical solutions of this application, and all such modifications and substitutions should be covered within the scope of the claims of this application.

Claims

1. A deep learning-based airborne hyperspectral image compression method, characterized in that, include: Acquire an airborne hyperspectral image data cube, the data cube comprising spatial and spectral dimensions; For each spectral band, calculate the spectral band reliability index and generate error boundary parameters for the spectral band; A reconstruction loop is established at the encoding end, and the reconstructed data is used as the prediction context input to the pre-trained deep learning prediction network to generate predicted values. The residual between the original data and the predicted values ​​is calculated. The pre-trained deep learning prediction network obtains network parameters in the offline training stage using a task-driven training method. In the training stage, a land cover classification task network is introduced and the category labels of the training samples are used as supervision information. The training objective consists of a bit rate estimation term, a task loss term, and a reconstruction constraint term, and is weighted and summed using preset non-negative weight coefficients. The residual is subjected to near-lossless quantization based on the error boundary parameter to generate a quantized residual; The quantization residual is entropy encoded to generate a compressed bitstream, and parameter information for error boundary generation, quantization rules, prediction context configuration, and model identification is written into the bitstream header.

2. The airborne hyperspectral image compression method based on deep learning according to claim 1, characterized in that: The spectral band reliability index is constructed from deterministic sources, namely, the estimated variance of the spectral band noise, the calibration uncertainty of the spectral band, and the spectral band anomaly rate. The spectral anomaly rate represents the proportion of bad elements and strip noise in the spectral band. The estimated value of the spectral noise variance and the spectral calibration uncertainty are obtained from the output of the airborne calibration link or from the statistics of the data cube at the airborne end.

3. The airborne hyperspectral image compression method based on deep learning according to claim 2, characterized in that: The error bound parameter is generated through a deterministic mapping from reliability to error bound. The mapping includes constraints on the lower and upper limits of the error bound. The reliability is then transformed by a preset proportional coefficient and a zero-prevention constant and truncated to the range between the lower and upper limits. The error bound parameters generated by the mapping are written into the code stream header in the form of spectral segment-level parameters, which are used by the decoding end to quantize and dequantize each spectral segment using a consistent error bound. The spectral band is divided into K spectral band groups, based on the numerical range boundary of the spectral band reliability index. For each spectral band group, the group reliability is calculated and a group error boundary parameter is generated. All spectral bands belonging to the same spectral band group are uniformly quantized and dequantized using the group error boundary parameter. The index mapping table of spectral bands to spectral band groups and the group error boundary parameter of each spectral band group are written into the code stream header.

4. The airborne hyperspectral image compression method based on deep learning according to claim 1, characterized in that: The near-lossless quantization employs a quantization step size generation rule that corresponds one-to-one with the error boundary parameters. This quantization step size generation rule satisfies the following formula: in, For spectral band error boundary parameters, Quantization step size for the spectral band; When using the band group error boundary parameter, bands belonging to the same band group share the same quantization step size; The encoding end performs scalar quantization on the residual according to the quantization step size to obtain the quantized residual. The decoding end performs dequantization according to the same quantization step size to reproduce and reconstruct the residual.

5. The airborne hyperspectral image compression method based on deep learning according to claim 1, characterized in that: The pre-trained deep learning prediction network uses the prediction context of the reconstruction loop at the encoding end to perform sample-by-sample prediction. The sample-by-sample prediction processes the spatial pixel position and spectral position sequentially according to a preset scanning order. The prediction context consists of the reconstructed pixel values ​​in the spatial neighborhood of the current sample to be predicted, the reconstructed spectral values ​​before the current sample to be predicted, and a context mask used to indicate the boundaries of the prediction context. After generating the quantization residual, the encoder immediately performs dequantization and superimposes it with the predicted value to obtain the reconstructed value of the current sample. The reconstructed value is then written into the reconstruction loop buffer for subsequent sample prediction.

6. The airborne hyperspectral image compression method based on deep learning according to claim 1, characterized in that: The parameter information written in the code stream header includes: spectral segment-level error boundary parameters or spectral segment group error boundary parameters, index mapping table from spectral segment to spectral segment group, lower and upper error boundaries, proportional coefficient and zero-prevention constant used for reliability-to-error boundary mapping, quantization step size generation rule identifier, model identifier of pre-trained deep learning prediction network, prediction context configuration parameters, and scan order identifier, wherein the prediction context configuration parameters include spatial neighborhood window size and the number of spectral reference segments; After reading the parameter information from the bitstream header, the decoding end performs the following decoding process: entropy decoding is performed on the entropy-encoded data to obtain the quantization residual; The quantization step size of each spectral segment is calculated based on the error boundary parameters and quantization step size generation rules in the code stream header, and inverse quantization is performed to obtain the reconstruction residual; At the decoding end, a reconstruction loop buffer consistent with that at the encoding end is established. A prediction context is generated according to the scanning order indicated by the code stream header and the prediction context configuration parameters. The prediction context is input into the pre-trained deep learning prediction network corresponding to the model identifier to generate prediction values. The prediction values ​​are superimposed with the reconstruction residuals to obtain reconstruction data. At the same time, the reconstruction data is written into the reconstruction loop buffer for subsequent sample prediction calls.

7. The airborne hyperspectral image compression method based on deep learning according to claim 1, characterized in that: The task loss term takes the form of cross-entropy and satisfies the following equation: in, This is the task loss value. For the number of categories, The first label for the category dimensional components, The first output result of the task network on the reconstructed data dimensional components, This is a logarithmic operation.

8. An airborne hyperspectral image compression system based on deep learning, characterized in that, include: The acquisition module is used to acquire an airborne hyperspectral image data cube, wherein the data cube contains spatial dimension and spectral dimension; The calculation module is used to calculate the reliability index of each spectral band and generate error boundary parameters for the spectral band. The prediction module is used to establish a reconstruction loop at the encoding end and use the reconstructed data as the prediction context input to the pre-trained deep learning prediction network to generate predicted values, and calculate the residual between the original data and the predicted values. The pre-trained deep learning prediction network obtains network parameters in the offline training stage using a task-driven training method. In the training stage, a land cover classification task network is introduced and the category labels of the training samples are used as supervision information. The training objective consists of a bit rate estimation term, a task loss term, and a reconstruction constraint term, and is weighted and summed using preset non-negative weight coefficients. The generation module is used to perform near-lossless quantization on the residual based on the error boundary parameter to generate quantized residual; The writing module is used to entropy encode the quantization residual to generate a compressed bitstream, and write parameter information for the decoding end to reproduce error boundary generation, quantization rules, prediction context configuration and model identification into the bitstream header.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the airborne hyperspectral image compression method based on deep learning as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the airborne hyperspectral image compression method based on deep learning as described in any one of claims 1 to 8.

Images